12.3 Approaches to multi-tasking

Two approaches to multi-tasking MOM have been experimented with at GFDL. They were:

• The fine grained approach to multi-tasking where parallelism is applied at the level of each nested do loop. This is also known as ``autotasking''. All processors work simultaneously on each nested do loop and there are many parallel regions.

• The coarse grained approach to multi-tasking where parallelism is applied at the level of latitude rows. For example, all work associated with solving equations for one latitude row is assigned to one processor. All work associated with solving the equations for another latitude row is assigned to a second processor ...and so forth. Then, all processors work simultaneously and independently. This implies that there is only one parallel region.

In a nutshell, the fine grained approach to multi-tasking gives about 85% parallelism. That translates into a speedup of a little less than 4 on 8 processors. Why is the parallel efficiency so low? The reason is basically due to not enough work within the many parallel regions and the fact that the range on the parallel loop ``j'' index is not the same for all loops. Increasing the model resolution can somewhat address the first problem but not the second. The result is that processors stand idle and the parallel performance degrades. This is clearly not a long term solution. The coarse grained approach is the best and yields a near linear speedup with the number of processors.